Research Papers

Artificial intelligence (AI) and Machine Learning (ML) have moved from research and pilot projects into everyday business operations, with generative AI accelerating adoption across processes, products, and services. This paper introduces the concept of Responsible AI for organizational practice, with a particular focus on small and medium-sized enterprises. It structures Responsible AI along four focal areas that are central for introducing and operating AI systems in a legally compliant, comprehensible, sustainable, and data-sovereign manner. First, it discusses the EU AI Act as a risk-based regulatory framework, including the distinction between provider and deployer roles and the resulting obligations such as risk assessment, documentation, transparency requirements, and AI literacy measures. Second, it addresses Explainable AI as a basis for transparency and trust, clarifying key notions such as transparency, interpretability, and explainability and summarizing practical approaches to make model behavior and decisions more understandable. Third, it covers Green AI, emphasizing that AI systems should be evaluated not only by performance but also by energy and resource consumption, and outlines levers such as model reuse, resource-efficient adaptation, continuous learning, model compression, and monitoring. Fourth, it examines local models (on-premise and edge) as an operating option that supports data protection, control, low latency, and strategic independence, including domain adaptation via fine-tuning and retrieval-augmented generation. The paper concludes with a consolidated set of next steps for establishing governance, documentation, secure operation, sustainability considerations, and an implementation roadmap.

Modern Integrated Development Environments (IDEs) increasingly leverage Large Language Models (LLMs) to provide advanced features like code autocomplete. While powerful, training these models on user-written code introduces significant privacy risks, making the models themselves a new type of data vulnerability. Malicious actors can exploit this by launching attacks to reconstruct sensitive training data or infer whether a specific code snippet was used for training. This paper investigates the use of Differential Privacy (DP) as a robust defense mechanism for training an LLM for Kotlin code completion. We fine-tune a \texttt{Mellum} model using DP and conduct a comprehensive evaluation of its privacy and utility. Our results demonstrate that DP provides a strong defense against Membership Inference Attacks (MIAs), reducing the attack's success rate close to a random guess (AUC from 0.901 to 0.606). Furthermore, we show that this privacy guarantee comes at a minimal cost to model performance, with the DP-trained model achieving utility scores comparable to its non-private counterpart, even when trained on 100x less data. Our findings suggest that DP is a practical and effective solution for building private and trustworthy AI-powered IDE features.

As AI systems evolve from isolated predictors into complex, heterogeneous ecosystems of foundation models and specialized adapters, distinguishing genuine behavioral novelty from functional redundancy becomes a critical governance challenge. Here, we introduce a statistical framework for auditing model uniqueness based on In-Silico Quasi-Experimental Design (ISQED). By enforcing matched interventions across models, we isolate intrinsic model identity and quantify uniqueness as the Peer-Inexpressible Residual (PIER), i.e. the component of a target's behavior strictly irreducible to any stochastic convex combination of its peers, with vanishing PIER characterizing when such a routing-based substitution becomes possible. We establish the theoretical foundations of ecosystem auditing through three key contributions. First, we prove a fundamental limitation of observational logs: uniqueness is mathematically non-identifiable without intervention control. Second, we derive a scaling law for active auditing, showing that our adaptive query protocol achieves minimax-optimal sample efficiency ($dσ^2γ^{-2}\log(Nd/δ)$). Third, we demonstrate that cooperative game-theoretic methods, such as Shapley values, fundamentally fail to detect redundancy. We implement this framework via the DISCO (Design-Integrated Synthetic Control) estimator and deploy it across diverse ecosystems, including computer vision models (ResNet/ConvNeXt/ViT), large language models (BERT/RoBERTa), and city-scale traffic forecasters. These results move trustworthy AI beyond explaining single models: they establish a principled, intervention-based science of auditing and governing heterogeneous model ecosystems.

Explainable AI (XAI) is essential for understanding machine learning (ML) decision-making and ensuring model trustworthiness in scientific applications. Prototype-based XAI methods offer an intrinsically interpretable alternative to post-hoc approaches which often yield inconsistent explanations. Prototype-based XAI methods make predictions based on the similarity between inputs and learned prototypes that represent typical characteristics of target classes. However, existing prototype-based models are primarily designed for standard RGB image data and are not optimized for the distinct, variable-specific channels commonly found in geoscientific image and raster datasets. In this study, we develop a prototype-based XAI approach tailored for multi-channel geospatial data, where each channel represents a distinct physical environmental variable or spectral channel. Our approach enables the model to identify separate, channel-specific prototypical characteristics sourced from multiple distinct training examples that inform how these features individually and in combination influence model prediction while achieving comparable performance to standard neural networks. We demonstrate this method through two geoscientific case studies: (1) classification of Madden Julian Oscillation phases using multi-variable climate data and (2) land-use classification from multispectral satellite imagery. This approach produces both local (instance-level) and global (model-level) explanations for providing insights into feature-relevance across channels. By explicitly incorporating channel-prototypes into the prediction process, we discuss how this approach enhances the transparency and trustworthiness of ML models for geoscientific learning tasks.

As generative AI increasingly mediates learning and decision-making, users often act effectively while struggling to interpret how system outcomes are produced. While Explainable Artificial Intelligence (XAI) research has primarily addressed this problem through transparency and visualization, less attention has been paid to how explanation is constructed through interaction. This paper examines digital games as explainable interfaces by analyzing how explanation can be configured as a playable process. Using Arknights as a case study, the paper conducts a qualitative close reading and interface analysis of the diegetic AI system PRTS, focusing on the implied player. The analysis shows that PRTS provides usable but unverifiable explanations: sufficient to initiate action, yet insufficient to stabilize causal understanding. Through incomplete information, delayed feedback, and narrative disruptions of trust, player agency is reorganized from direct control toward interpretive and abductive reasoning. The paper conceptualizes this mode as explanatory agency and discusses its implications for XAI-oriented interface design.

Large language models (LLMs) are increasingly deployed in high-stakes domains, where rare but severe failures can result in irreversible harm. However, prevailing evaluation benchmarks often reduce complex social risk to mean-centered scalar scores, thereby obscuring distributional structure, cross-dimensional interactions, and worst-case behavior. This paper introduces Social Harm Analysis via Risk Profiles (SHARP), a framework for multidimensional, distribution-aware evaluation of social harm. SHARP models harm as a multivariate random variable and integrates explicit decomposition into bias, fairness, ethics, and epistemic reliability with a union-of-failures aggregation reparameterized as additive cumulative log-risk. The framework further employs risk-sensitive distributional statistics, with Conditional Value at Risk (CVaR95) as a primary metric, to characterize worst-case model behavior. Application of SHARP to eleven frontier LLMs, evaluated on a fixed corpus of n=901 socially sensitive prompts, reveals that models with similar average risk can exhibit more than twofold differences in tail exposure and volatility. Across models, dimension-wise marginal tail behavior varies systematically across harm dimensions, with bias exhibiting the strongest tail severities, epistemic and fairness risks occupying intermediate regimes, and ethical misalignment consistently lower; together, these patterns reveal heterogeneous, model-dependent failure structures that scalar benchmarks conflate. These findings indicate that responsible evaluation and governance of LLMs require moving beyond scalar averages toward multidimensional, tail-sensitive risk profiling.

Endotracheal suctioning (ES) is an invasive yet essential clinical procedure that requires a high degree of skill to minimize patient risk - particularly in home care and educational settings, where consistent supervision may be limited. Despite its critical importance, automated recognition and feedback systems for ES training remain underexplored. To address this gap, this study proposes a unified, LLM-centered framework for video-based activity recognition benchmarked against conventional machine learning and deep learning approaches, and a pilot study on feedback generation. Within this framework, the Large Language Model (LLM) serves as the central reasoning module, performing both spatiotemporal activity recognition and explainable decision analysis from video data. Furthermore, the LLM is capable of verbalizing feedback in natural language, thereby translating complex technical insights into accessible, human-understandable guidance for trainees. Experimental results demonstrate that the proposed LLM-based approach outperforms baseline models, achieving an improvement of approximately 15-20\% in both accuracy and F1 score. Beyond recognition, the framework incorporates a pilot student-support module built upon anomaly detection and explainable AI (XAI) principles, which provides automated, interpretable feedback highlighting correct actions and suggesting targeted improvements. Collectively, these contributions establish a scalable, interpretable, and data-driven foundation for advancing nursing education, enhancing training efficiency, and ultimately improving patient safety.

Recent work identifies a stated-revealed (SvR) preference gap in language models (LMs): a mismatch between the values models endorse and the choices they make in context. Existing evaluations rely heavily on binary forced-choice prompting, which entangles genuine preferences with artifacts of the elicitation protocol. We systematically study how elicitation protocols affect SvR correlation across 24 LMs. Allowing neutrality and abstention during stated preference elicitation allows us to exclude weak signals, substantially improving Spearman's rank correlation ($ρ$) between volunteered stated preferences and forced-choice revealed preferences. However, further allowing abstention in revealed preferences drives $ρ$ to near-zero or negative values due to high neutrality rates. Finally, we find that system prompt steering using stated preferences during revealed preference elicitation does not reliably improve SvR correlation on AIRiskDilemmas. Together, our results show that SvR correlation is highly protocol-dependent and that preference elicitation requires methods that account for indeterminate preferences.

Machine learning models in high-stakes applications, such as recidivism prediction and automated personnel selection, often exhibit systematic performance disparities across sensitive subpopulations, raising critical concerns regarding algorithmic bias. Fairness auditing addresses these risks through two primary functions: certification, which verifies adherence to fairness constraints; and flagging, which isolates specific demographic groups experiencing disparate treatment. However, existing auditing techniques are frequently limited by restrictive distributional assumptions or prohibitive computational overhead. We propose a novel empirical likelihood-based (EL) framework that constructs robust statistical measures for model performance disparities. Unlike traditional methods, our approach is non-parametric; the proposed disparity statistics follow asymptotically chi-square or mixed chi-square distributions, ensuring valid inference without assuming underlying data distributions. This framework uses a constrained optimization profile that admits stable numerical solutions, facilitating both large-scale certification and efficient subpopulation discovery. Empirically, the EL methods outperform bootstrap-based approaches, yielding coverage rates closer to nominal levels while reducing computational latency by several orders of magnitude. We demonstrate the practical utility of this framework on the COMPAS dataset, where it successfully flags intersectional biases, specifically identifying a significantly higher positive prediction rate for African-American males under 25 and a systemic under-prediction for Caucasian females relative to the population mean.

While explainable artificial intelligence (XAI) for large language models (LLMs) remains an evolving field with many unresolved questions, increasing regulatory pressures have spurred interest in its role in ensuring transparency, accountability, and privacy-preserving machine unlearning. Despite recent advances in XAI have provided some insights, the specific role of attention layers in transformer based LLMs remains underexplored. This study investigates the memory mechanisms instantiated by attention layers, drawing on prior research in psychology and computational psycholinguistics that links Transformer attention to cue based retrieval in human memory. In this view, queries encode the retrieval context, keys index candidate memory traces, attention weights quantify cue trace similarity, and values carry the encoded content, jointly enabling the construction of a context representation that precedes and facilitates memory retrieval. Guided by the Encoding Specificity Principle, we hypothesize that the cues used in the initial stage of retrieval are instantiated as keywords. We provide converging evidence for this keywords-as-cues hypothesis. In addition, we isolate neurons within attention layers whose activations selectively encode and facilitate the retrieval of context-defining keywords. Consequently, these keywords can be extracted from identified neurons and further contribute to downstream applications such as unlearning.

A core assumption of Explainable AI (XAI) is that explanations are useful to users -- that is, users will do something with the explanations. Prior work, however, does not clearly connect the information provided in explanations to user actions to evaluate effectiveness. In this paper, we articulate this connection. We conducted a formative study through 14 interviews with end users in education and medicine. We contribute a catalog of information and associated actions. Our catalog maps 12 categories of information that participants described relying on to take 60 different actions. We show how AI Creators can use the catalog's specificity and breadth to articulate how they expect information in their explanations to lead to user actions and test their assumptions. We use an exemplar XAI system to illustrate this approach. We conclude by discussing how our catalog expands the design space for XAI systems to support actionability.

AI systems are increasingly used in high-stakes domains such as credit rating, where fairness concerns are critical. Existing fairness assessments are typically conducted by AI experts or regulators using predefined protected attributes and metrics, which often fail to capture the diversity and nuance of fairness notions held by the individuals who are affected by these systems' decisions, such as decision subjects. Recent work has therefore called for involving affected individuals in fairness assessment, yet little empirical evidence exists on how they create their own fairness criteria or what kinds of criteria they produce - knowledge that could not only inform experts' fairness evaluation and mitigation, but also guide the design of AI assessment tools. We address this gap through a qualitative user study with 18 participants in a credit rating scenario. Participants first articulated their fairness notions in their own words. Then, participants turned them into concrete quantified and operationalized fairness criteria, through an interactive prototype we designed. Our findings provide empirical evidence of the process through which people's fairness notions emerge via grounding in model features, and uncover a diverse set of individuals' custom-defined criteria for both outcome and procedural fairness. We provide design implications for processes and tools that support more inclusive and value-sensitive AI fairness assessment.